SocialWise: LLM-Agentic Conversation Therapy for Individuals with Autism Spectrum Disorder to Enhance Communication Skills

Based on the paper provided, here is a detailed technical summary of SocialWise, an LLM-AGentic Conversation Therapy system designed for individuals with Autism Spectrum Disorder (ASD).

1. Problem Statement

The paper identifies a critical gap in the support ecosystem for the estimated 75 million people worldwide with ASD, particularly regarding the practice of everyday conversational skills.

• Scalability vs. Efficacy: Low-cost interventions (e.g., story reading) yield limited improvements, while effective therapies (e.g., in-person role-play with specialists) are constrained by high costs, geographical limitations, and provider scarcity.
• Limitations of Existing Tech: Current digital aids, such as Augmentative-and-Alternative Communication (AAC) tools and Augmented Reality (AR) systems, often still require therapist supervision and lack adaptive, personalized feedback.
• The Gap: There is a lack of a scalable, therapist-free solution that offers realistic role-play scenarios coupled with evidence-based, personalized feedback tailored to ASD communication goals.

2. Methodology & System Architecture

SocialWise is a browser-based application designed to bridge this gap by combining Large Language Models (LLMs) with a therapeutic knowledge base.

• Core Technology Stack:
  • Frontend: Built using Streamlit for rapid deployment and web accessibility.
  • LLM Dialogue: Utilizes GPT-4o-mini for generating natural, context-aware conversational responses.
  • Multimodal Interaction: Integrates OpenAI Whisper for Speech-to-Text (STT) and OpenAI TTS for Text-to-Speech, allowing users to interact via voice or text.
  • Retrieval-Augmented Generation (RAG): Employs LangChain and ChromaDB (a dense vector retriever) to ground the system's feedback in a curated database of therapeutic documents and ASD-specific guidelines.
• Workflow:
  1. Scenario Selection: Users choose a preset scenario (e.g., "Ordering food," "Joining a group") or input a custom situation.
  2. Role-Play: The user engages in a live conversation with an LLM agent.
  3. Analysis & Feedback: Upon request, a feedback agent analyzes the user's responses against the RAG-retrieved therapeutic guidelines.
  4. Output: The system provides structured, instant feedback highlighting strengths, areas for improvement, and alternative phrasing.
• Safety & Ethics Mechanisms:
  • Privacy: All conversation logs are stored only in the browser's memory and are flushed upon page refresh; no audio or identifiers are stored server-side.
  • Bias & Hallucination Control: The system uses Chain-of-Thought (CoT) bias-mitigation prompts, strict instruction prompts, and RAG grounding to minimize factual drift and hallucinations.

3. Key Contributions

The paper highlights three primary contributions to the field of Human-Computer Interaction (HCI) and assistive technology:

1. Therapist-Free Accessibility: A fully functional, multimodal web interface that delivers therapist-style role-play and feedback without requiring specialized hardware or professional supervision.
2. Evidence-Based Feedback Agent: An LLM-driven feedback mechanism that is not generic but is specifically grounded in curated therapeutic documents, ensuring alignment with evidence-based ASD therapy guidelines.
3. Preliminary Validation: The release of preliminary user data (N=34) demonstrating high perceived helpfulness and user willingness to recommend the tool.

4. Results (Preliminary User Feedback)

A remote pilot study was conducted with 34 autistic adults (ages 18–34). The results indicated strong perceived value:

• Helpfulness Rating: Participants rated the overall helpfulness at 4.15 / 5.
• Recommendation Rate: 100% of participants stated they would recommend the tool to a peer.
• Key Benefits Cited:
  • Improved turn-taking (67.6%).
  • Enhanced problem-solving in social conflicts (52.9%).
  • Reduced anxiety during real-world conversations (50.0%).
• User Feedback: Participants expressed a desire for gamified rewards and more advanced scenarios, indicating a strong appetite for deeper practice.

5. Significance and Future Work

Significance:
SocialWise demonstrates how recent advances in LLMs and RAG can democratize access to high-quality communication coaching. By removing the barriers of cost and geography, it offers a scalable alternative to traditional speech therapy, potentially improving employment outcomes and social integration for the ASD community.

Future Work:
The authors outline several directions for future development:

• Clinical Validation: Conducting formal clinical trials to validate efficacy beyond user perception.
• Content Expansion: Broadening the library of scenarios to cover diverse social contexts.
• Platform Expansion: Developing a dedicated mobile application.
• Progress Tracking: Implementing features to track user progress over time.
• Model Refinement: Fine-tuning the underlying models based on community feedback to further reduce bias.

In conclusion, SocialWise represents a significant step toward making evidence-based, on-demand social skills training accessible to individuals with ASD through a secure, privacy-focused, and technically robust web application.